__GOOGLE_SITE_VERIFICATION_META__ ThumbGate — Warn Claude Code before it repeats a mistake __GA_BOOTSTRAP__
PreToolUse gate for Claude Code

Thumbs-down a Claude Code mistake once, and ThumbGate warns the next time it tries that command. After three thumbs-downs on the same command, strict mode blocks it.

Run Claude Code with bypass permissions and ThumbGate checks every call: secret leaks and catastrophic commands are denied, other flagged actions warn, strict mode blocks them.

Who it's for: Claude Code users whose agent has already broken something once. The local CLI is free.

Get Pro — $__PRO_PRICE_DOLLARS__/month
Real 58-second run with THUMBGATE_STRICT_ENFORCEMENT=1: three thumbs-downs, then the fourth attempt is denied. A reworded delete is not caught. How it was recorded.
$ echo '{"tool_name":"Bash","tool_input":{"command":"git push --force origin main"}}' | thumbgate gate-check
{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny",
 "permissionDecisionReason":"[GATE:force-push] Force push blocked. This is destructive and irreversible."}}

$ ... "git reset --hard HEAD~5"
 "permissionDecisionReason":"[GATE:git-reset-hard] git reset --hard is blocked because it can destroy uncommitted work. Stash or use a non-destructive restore path ..."

$ ... "rm -rf ~/"
 "permissionDecisionReason":"[GATE:rm-rf-home-or-root] Broad rm -rf against root or home is blocked. Use a narrow, reviewed path ..."

Recorded from thumbgate gate-check v1.40.3. Claude reads the permissionDecisionReason and changes course. These three deny by default. Hard deny for every other flagged action is strict mode, one env var: THUMBGATE_STRICT_ENFORCEMENT=1.

1 · Install the plugin inside Claude Code

/plugin marketplace add IgorGanapolsky/ThumbGate
/plugin install thumbgate@thumbgate-marketplace

2 · Set up the status line in your terminal, no manual config

npx thumbgate init --agent claude 

Step 2 writes the hooks and the status line into your Claude Code settings. Restart Claude Code and the line below appears.

  • 0👍 2👎 this project Live count of the thumbs you gave in the project you are in. Automated and imported feedback is not counted.
  • Dashboard One click opens your local dashboard.
  • Latest success … The most recent lesson ThumbGate captured.
Real capture from a Claude Code session on 16 Nov 2026 (ThumbGate v1.40.3, Pro). Your counts and lesson will differ.
How it works

Catch it. Signal it. Enforce it.

  1. 1 · Catch Claude tries something you rejected before, like npm publish from the wrong branch, or mocking the database in an end-to-end test.
  2. 2 · Signal One thumbs-down with the reason: npx thumbgate capture --feedback=down --context="mocked the DB in an e2e test"
  3. 3 · Enforce A local PreToolUse gate checks every future tool call in every Claude Code session in that project. A match warns Claude with your reason; in strict mode, after three thumbs-downs, it is denied.

What init adds to your Claude Code settings

{
  "statusLine": { "type": "command", "command": "node \"<path to thumbgate>/bin/cli.js\" statusline-render" },
  "hooks": {
    "PreToolUse": [
      { "matcher": ".*",
        "hooks": [ { "type": "command", "command": "node \"<path to thumbgate>/bin/cli.js\" gate-check" } ] }
    ]
  }
}

Written by npx thumbgate init --agent claude into a fresh Claude Code home; the path points at the installed package. They are ordinary Claude Code hooks, so they show up in /hooks.

Compared with doing it yourself

Your own hooks vs ThumbGate.

Hand-written PreToolUse hooks and settings.json deny rules versus ThumbGate
Own hooks + deny rulesThumbGate
Where a rule comes fromYou write the pattern by handOne thumbs-down with a reason becomes a warning rule
Across sessionsWhatever you saved in settingsLearned rules live on disk and load every session
What Claude is toldWhatever message you wroteYour own reason is shown to Claude in the warning
Visible in /hooksYesYes
Claude Code · local gate

A firewall for AI coding agents, built for Claude Code. Other agents are listed under Also works with.

Dashboard: npx thumbgate dashboard --open · /thumbgate-dashboard · thumbgate-dashboard · demo · strict-mode deny example ↓

Why Pro is worth $__PRO_PRICE_DOLLARS__

Free blocks force-push, git reset --hard, git clean -f, rm -rf on home or root and DROP TABLE by default. Pro is about memory.

  • Every learned rule stays active. Free keeps your first 3 learned rules active. Pro keeps all of them.
  • Recall. Ask what Claude got wrong before, and search past lessons. Both are Pro only.
  • Coming, not shipped yet: team-shared rules, lessons synced across machines, audit history.

Get Pro — $__PRO_PRICE_DOLLARS__/month

Thumbs teach. The gate enforces.

Without ThumbGate the same AI agent mistake repeats. With ThumbGate one thumbs-down becomes a local pre-action gate that can refresh or expire.
Agent proposes a tool, ThumbGate decides ALLOW WARN or DENY, thumbs feedback becomes a rule that closes the loop.
Self-improving product loop

Pre-action checks—and the systems that refine them.

Corrections become local lessons; repeated negatives become gates; the next action is checked before execution. Click a step for the under-the-hood demo.

Self-improving ThumbGate loop: thumbs up or thumbs down become a local lesson, then a prevention rule, then ALLOW WARN or DENY on the next tool call. The loop closes under your control without retraining the model.
Is it really self-improving? Yes — the control layer, not the LLM. 👍/👎 → local lesson → ranked rule → next tool call gated. No model retrain.

Each reviewed outcome closes the loop—under your control. Lessons are re-ranked per action, repeated failures can promote into gates, and stale auto-promoted gates expire. The firewall improves from explicit feedback without retraining the model or silently rewriting policy.

What "stop" honestly means

Hard deny where configured. Honest warning everywhere else.

proposed  git push --force origin main
mode      strict enforcement
rule      protect-main
source    core protection · strict mode
decision  DENY
reason    force-push to protected branch
next      decision recorded before execution

Why not a static hook? Prompts compact away; static bash hooks don't learn from 👍/👎. Detected secret exfiltration and gate-process bypass attempts are denied by default. Catastrophic commands (force-push, reset --hard, rm -rf on home or root, DROP TABLE) are denied by default. Other matching destructive actions warn by default and deny in strict mode. This is a managed strict-mode example, not a claim that every free install blocks every risky command automatically.

Read the test-backed verification evidence →

White paper · Scorecard · Architecture · How we use it

Before you buy

Five questions, answered plainly.

The MIT local CLI captures feedback and runs configured gates. Install the Claude Code plugin, then run npx thumbgate init --agent claude.
Pro adds memory: every learned rule stays active, plus recall and lesson search. Strict blocking of catastrophic commands stays free. Compare current subscription details at /pricing.
Force-push, git reset --hard, git clean -f, rm -rf on home or root and DROP TABLE are denied by default; hard deny for every other matching action needs strict mode (THUMBGATE_STRICT_ENFORCEMENT=1), otherwise they warn and are logged. Detected secret exfiltration and gate-process bypass attempts are denied by default. It does not retrain model weights, does not rewrite policy silently, and does not contain the blast radius, which is a sandbox job. Only the agents in Also works with are supported, and advisory ones do not guarantee blocking.
Thumbs teach. The gate enforces. Feedback refines the control layer, not the LLM. Reviewed outcomes can become prevention rules under your control.
Claude Code enforces gates through PreToolUse hooks. Other agents are listed under Also works with; advisory integrations do not guarantee execution blocking.
Start free today

Get warned before your agent repeats last week's mistake.

Run npx thumbgate init (MIT), or choose self-serve Pro.

Install free Get Pro
Also works with

ThumbGate also installs for Cursor, Codex, Gemini CLI, Cline, OpenCode, Amp, Grok and Hermes. Claude Code is the first-class path.

Enforcement by agent and mode

Per-agent enforcement and default block vs warn
AgentEnforcementDefault (secrets)Default (destructive)Strict mode
Claude CodeBlocks before tool runsDENYDENY catastrophic, WARN otherDENY
Cursor, Cline, OpenCodeAdvisoryDENYWARNDENY
AmpFeedback onlyLoggedLoggedLogged